Crypto Custody ⚡ Google MPC for Bank Private Keys
Learn how Google Cloud crypto custody uses MPC to secure private key security across environments—understand risks, benefits, and what institutional crypto teams must assess.

Google's MPC Custody Tool: What Banks Need to Know
For a bank holding digital assets, custody isn't just about storing private keys. It's about who can approve a transfer, how many pieces have to agree, and what happens when something goes wrong - an employee account, a device, or a third-party service. One exposed signing credential can turn an operational mistake into a loss you can't claw back.
Google is previewing a custody system that tries to spread signing authority across separate environments. It's based on Google Multi-Party Compute (MPC), which splits a key into shares and requires a threshold number of them to approve a transaction. Google says policy controls run inside Google Confidential Space and that it can't access key shares in unencrypted form. Several financial institutions are testing the system ahead of planned commercial availability in early 2027.
This could give banks another option for reducing reliance on a single device or provider. But distributing key shares doesn't magically erase risk. Below is how Google's system is reported to work, what MPC is meant to fix, how it compares with existing custody setups, and what institutions should check before adopting it.
What is Google Multi-Party Compute?
Google Multi-Party Compute is a crypto custody approach built on multi-party computation. Instead of keeping one complete signing key in a single place, MPC splits the key into separate shares. A transaction can be signed only when enough of those shares participate, based on a threshold the institution sets.
For instance, a bank might choose a two-of-three setup: one share held on bank-controlled hardware, another in Google Cloud, and a third with a separate provider. Any two shares could authorize a signature, but compromising one environment by itself wouldn't be enough. Larger institutions might use a five-of-seven threshold to spread authority across more teams, devices, or locations.
The point of the design is that the full private key isn't assembled in one place during signing. The participating systems use a protocol to generate a signature without exposing the entire key. That can limit how much damage a breach in one environment causes - but each share and the systems coordinating signing still have to be secured.
Google says a policy engine running inside Confidential Space checks transactions against institutional rules before signing. Those rules could include transfer limits, permitted destinations, and required approvals. Google also says it can't access the key shares in unencrypted form. The system is in private preview, and reporting says Alphabet has begun using it to secure its own assets. Google hasn't published pricing, and it expects commercial availability in early 2027.
MPC shouldn't be mixed up with a blockchain-native multisignature feature. Native multisig typically requires multiple distinct on-chain signatures and may be visible in the transaction structure. MPC can coordinate participants to produce a signature that may appear to the blockchain like a conventional single signature. That difference matters for auditing, operational visibility, and recovery planning - banks need to understand not just how authority is split, but how approvals and the resulting transactions get reviewed afterward.
How does MPC custody change the security model?
MPC shifts where signing authority "lives." In many traditional setups, a key is protected by hardware security modules, offline storage, or tightly controlled access. Those controls can be strong, but the key - or the ability to use it - may still rely on a small set of systems or people. If an attacker gets the right credential or compromises the signing environment, there may not be any meaningful barrier left.
With threshold MPC, the institution sets a rule requiring multiple key shares to cooperate. The intent is to turn a single point of failure into something closer to a coordination problem. If an attacker compromises one machine or provider, they still need additional shares and have to satisfy the relevant controls. In practice, splitting shares across organizations can also reduce dependency on one vendor - assuming the separation is real and not just theoretical.
The catch is that "distributed" doesn't automatically mean "independent." If several shares are administered by the same team, exposed through the same identity system, or deployed into the same cloud account, a shared weakness could affect them all at once. Banks should map the full chain: how shares are stored, how identity and access are handled, how transaction policies are enforced, how approvals are routed, how monitoring works, and how backup and recovery are done. A threshold only helps if the counted parties aren't sharing avoidable vulnerabilities.
The broader industry is still focused on signing controls for a reason. Chainalysis attributed 43.8% of cryptocurrency stolen in 2024 to compromised private keys. DeFiLlama data cited in reporting suggested that about 40% of the $16.69 billion lost to hacks through June came from key compromises. These numbers come from different datasets and time windows, so you can't treat them as a clean apples-to-apples comparison. Still, they point to a persistent operational risk: when signing authority is taken, attackers can move assets quickly and irreversibly.
Even with MPC, not every transaction becomes "safe." A compromised employee account, a manipulated approval workflow, or a malicious change to a destination rule could still result in an authorized-looking transfer. The practical improvement is narrower: it's harder for an attacker with one stolen credential or one breached environment to reach full signing authority. That's a real reduction in risk, but it doesn't replace controls around people, policies, and incident response.
Why are banks evaluating Google's custody tool?
Banks look at custody systems because they need operational controls that fit governance, audit, and risk processes - not just a way to store tokens. A custody platform has to answer basic questions: who can initiate a transfer, who approves it, which assets and destinations are allowed, how exceptions are handled, and how the institution can investigate a disputed transaction later.
Google's pitch combines threshold signing with policy enforcement and its cloud infrastructure. A bank might require two separate parties to approve a transfer while applying limits and destination rules before a signature is produced. The ability to split shares between the bank and different providers may appeal to institutions that don't want all signing authority to sit under one vendor. Google also argues that its security expertise and existing relationships with banks help it stand out in the market.
Google is entering a crowded field. Fireblocks, Fordefi, and BitGo already provide digital-asset custody and wallet infrastructure. Banks have also used specialist providers - BNP Paribas has custody work with Metaco, and BitGo has been involved in multi-institution vault arrangements. The pattern is familiar: institutions often combine external crypto infrastructure with internal controls and third-party assurance rather than building everything in-house.
So Google's technology may be one option in a vendor and architecture decision, not a replacement for the entire custody stack. The comparison shouldn't be "MPC versus no MPC." It's how the provider handles share independence, transaction policies, supported assets and networks, service continuity, audit evidence, recovery, and liability. A bank also needs to understand which parts it controls directly and which depend on Google or another service provider.
There's also a commercial and strategic angle. Some institutions may see a major cloud provider as a credible infrastructure partner. Others may be wary of concentrating operational dependencies with a company that already hosts workloads they rely on. A neutral supplier can help when banks don't want to rely on a direct competitor - but neutrality doesn't eliminate vendor lock-in or concentration risk. The assessment has to be specific: what does the bank gain, what dependency does it add, and is there a tested fallback if the provider changes or becomes unavailable?
What should institutions assess before adopting it?
A private preview is a chance to test whether the workflow fits your operations - not proof the system is ready for production. Before adopting Google's tool (or any institutional custody platform), banks should evaluate the end-to-end process: from transaction request to recovery. The cryptographic design matters, but so do the people and systems around it.
A practical assessment should cover at least these areas:
- Share independence: Confirm where each share is stored, who can access it, and whether separate shares depend on the same identity provider, administrators, or network controls.
- Threshold and approval design: Test whether the chosen quorum matches staffing, business hours, and emergency procedures. A threshold that's too strict can slow legitimate transfers; one that's too permissive weakens separation.
- Policy enforcement: Check how limits, destination allowlists, and approval rules are set, changed, and audited. Also confirm who can override policy and what evidence that leaves behind.
- Recovery and continuity: Run scenarios for a lost share, an unavailable provider, staff departure, and a regional outage. Recovery shouldn't require an undocumented exception or reintroduce the same concentration the design is trying to avoid.
- Audit and incident response: Determine what transaction records, approvals, and security events the institution can review, and how quickly it can suspend signing after a suspected compromise.
- Portability and exit: Ask how assets and controls can be migrated if the contract ends, the product changes, or the provider is unavailable. Custody shouldn't make exit dependent on a vendor's goodwill.
Banks should also reflect MPC in their threat models instead of assuming it solves everything. Google has framed AI-enabled fraud as part of the broader security context. Chainalysis reported that scams linked to AI tool vendors averaged $3.2 million per operation in 2025, compared with $719,000 for other scams. Those figures show an association in the report; they don't prove AI caused the difference. MPC may make stolen credentials less useful for unauthorized transfers, but it can't determine whether a legitimate approver was tricked into approving a fraudulent request.
Quantum computing is another, longer-term concern. Google researchers have estimated that a sufficiently powerful future quantum machine could derive a bitcoin private key from an exposed public key in about nine minutes; roughly 6.9 million bitcoin have exposed public keys. That's a forecast about future capability, not evidence that current machines can break Bitcoin's signatures. Google says its architecture can adopt post-quantum signatures later without forcing institutions to rebuild their systems, but it hasn't provided a timetable. Banks should treat migration readiness as a design question and avoid reading it as an immediate guarantee.
The core evaluation question is straightforward: does this system reduce the institution's real concentration risk for signing, without creating operational dependencies it can't manage? You only get the answer by testing the full custody workflow - including failure cases - rather than taking a provider's cryptographic description at face value.
Conclusion
Google's MPC custody approach targets a real operational weakness in digital-asset custody: when signing authority is concentrated in one key, one device, or one provider, it can become a single point of failure. Splitting authority across environments and layering in transaction policies can make unauthorized transfers harder - especially when an attacker has only one credential or foothold.
But the security outcome depends on the full system. Shares need to be meaningfully independent, policies have to be governed and audited, and recovery has to work when things go sideways. Banks should compare Google's design against existing custody providers using those operational requirements - not assume that a new cryptographic label automatically fixes vendor concentration, human error, or fraud.
For institutions considering the private preview, the next step is a threat-model review and failure simulation. Test what happens when a share, provider, approval account, or policy control is unavailable or compromised. If the system reduces the risks that matter without making recovery and oversight harder, it may earn a place in the custody stack. If not, the technology name won't save it.